Skip to main content
. 2021 Jul 23;12:4506. doi: 10.1038/s41467-021-24082-z

Table 2.

PRS prediction accuracy for the AD case-control dataset using different p-value thresholds and methods to model APOE.

pT PRS.full PRS.no.APOE PRS.AD
N SNPs AUC (%) R2 OR
(95% CI)
N SNPs AUC (%) R2 OR
(95% CI)
AUC
(%)
R2 OR
(95% CI)
APOE(ε2 + ε4) 2 70.0 0.18 2.2 (1.8,2.7) 70.0 0.18 2.2 (1.8, 2.7)
5e-8 65 69.8 0.16 2.2 (1.8, 2.7) 17 55.7 0.02 1.2 (1.0, 1.5) 71.4 0.19 2.4 (2.0, 3.0)
1e-5 (ORS) 126 69.4 0.16 2.2 (1.8, 2.7) 66 56.7 0.02 1.2 (1.1, 1.5) 72.0 0.20 2.4 (2.0, 3.0)
0.1 68,681 64.9 0.09 1.8 (1.5, 2.2) 68,516 61.3 0.06 1.6 (1.3, 1.9) 74.1 0.24 2.8 (2.2, 3.4)
0.5 203,950 62.6 0.07 1.7 (1.4, 2.0) 203,710 60.5 0.05 1.5 (1.3, 1.8) 73.7 0.23 2.7 (2.2, 3.4)

Legend: PRSs were calculated on a case-control cohort (271 clinically defined AD cases and 278 cognitively normal controls) using Kunkle et al. (2019) summary statistics for pT ≤ 5e-8, 1e-5, 0.1, 0.5 LD-pruned SNPs and APOE(ε2 + ε4). The number of SNPs (NSNPs) in each risk score are reported. Three PRS models were considered: PRS.full calculated on the full summary statistics; PRS.no.APOE where the APOE region was excluded (chr19:44.4–46.5 Mb); PRS.AD which is calculated as a weighted sum of PRS.no.APOE and APOE(ε2 + ε4), where APOE effects were weighted with effect sizes (B(ε2) = −0.47 and B(ε4) = 1.12) as in Kunkle et al (2019). The number of SNPs for PRS.AD models is always two more than for PRS.no.APOE. Prediction was estimated in terms of AUC, R2 and OR with 95% Confidence Intervals (CI).